Table 1. Overview of user inputs for the three simulated scenarios.

User inputs	Scenario 1	Scenario 2	Scenario 3
Candidate model space	RW	RW, KRW	RW(V), RWPH(V), RWPH(α), RWPH(V + α)
Analysis procedure	Maximum likelihood parameter estimation	Model selection using BIC	Model selection using BIC
Utility function	Absolute error of learning rate (α) estimate	Model selection accuracy	Model selection accuracy
Analysis prior	Uniform over α	Uniform over models and parameters	Uniform over models and parameters
Reference design	Acquisition followed by extinction of equal length	Backward blocking	Reversal learning
Evaluation prior	Point priors on low, middle and high α (LA, MA, HA)	Uniform over models with point priors on parameters (from [39])	Uniform over models with point priors on parameters (from [31])
Optimized experiment structure	One cue with periodically varying contingency	Two stages with three cues (A, B, AB) and stage-wise contingencies	Two stages with two cues (A, B) and stage-wise contingencies
Design space	Two contingencies (P1, P2) and the period (T) of their switching (3 variables)	For each stage s and cue X: Ps(X), Ps(US|X) (10 non-redundant variables)	For each stage s and cue X: Ps(X), Ps(US|X) (6 non-redundant variables)
Design priors	Either a point prior over α coinciding with evaluation prior (PA) or a vague prior (VA)	Uniform over models with either a point (PP) or vague (VP) prior over parameters	Uniform over models with either a point (PP) or vague (VP) prior over parameters

The role of user inputs is clarified by Fig 1. See the Methods section for details. Analysis priors are used in fitting models, design priors are used to simulate data when optimizing designs, and evaluation priors are used to simulate data when evaluating both reference and optimized designs.